Methodology · data as of Oct 2, 2026

Where the numbers come from.

Every morning a scheduled job fetches each source, validates it against a schema, and rejects suspicious changes, such as a dataset suddenly losing rows. Numbers that pass publish automatically; proposed timeline entries wait for review. If a source fails, the site keeps its last good data.

Status labels

• Verified

  Measured or published by an independent or primary source.

• Projection

  A model output or forward-looking estimate, labelled as such.

• Claim

  Reported by a lab or the press but not yet independently checked.

Definitions

Capability level (ECI)

Epoch AI's Capabilities Index combines many benchmarks into one score per model, anchored at Claude 3.5 Sonnet = 130 and GPT-5 = 150. A model reaches level X when its point estimate is at or above X. Epoch re-fits the index as results arrive, so scores can shift slightly; every sync stores the latest fit.

Price

List price in US dollars per million tokens, blended 3:1 input to output, from the providers' own published prices (via the LiteLLM price map). Only model labs, cloud platforms, and inference hosts count; aggregators and gateways are excluded. Per-token prices understate the cost of reasoning models, which use more tokens per answer.

Launch price and debut

A level debuts with the first model to reach it. Its launch price is the cheapest launch price among models reaching the level within 30 days, so a premium tier does not set the bar alone. Launch prices before tracking began come from archived copies of each lab's pricing page or launch post; after that, from the first price the sync records within 30 days of release.

Open-weight lag

Months from a level's debut to the first model with downloadable weights at or above it. The headline lag uses the most recent closed-weight record that an open-weight model has matched.

Task horizon

METR's measure of the length of task, in human expert time, that a model completes with 50% reliability on software, ML, and cybersecurity tasks. METR flags estimates above 16 hours as unreliable with its current task suite; they are shown as provisional.

Productivity

For the US, output per hour in the nonfarm business sector (BLS, quarterly). For cross-country comparison, GDP per person employed in constant PPP dollars (World Bank, annual). The two are not the same measure; each is labelled where shown.

Unemployment

US rates come from the BLS (youth: ages 16–24). Other countries use OECD harmonised rates on ILO definitions (youth: 15–24); the EU aggregate comes from Eurostat. Labour data move with interest rates and the business cycle, so the site shows them as context, not as evidence that AI caused a change.

AI-exposed job postings

Indeed's postings by occupation category, grouped into AI-exposed desk work and less-exposed hands-on work, each rebased so November 2022 (ChatGPT's launch month) = 100. A wider gap is consistent with AI reducing demand for desk work, but postings also respond to rates, sector slumps, and the business cycle.

Occupations and tasks

Occupations and their tasks come from O*NET; employment and wages from the BLS's Occupational Employment and Wage Statistics (US only). Task-level AI use comes from the Anthropic Economic Index, which reflects one AI system's usage, so it shows where AI is used, not everything AI could do.

Task-length projection

The calculator starts from the longest horizon METR has measured at the chosen reliability and extends it at METR's fitted doubling times: the post-2023 estimate with its confidence bounds, and the slower 2019-onward average. It is arithmetic on a trend line, not a forecast, and assumes the 80% horizon doubles at the same pace as the 50% one.

Long-run income and growth

Real GDP per person in 2011 international dollars from the Maddison Project Database. Growth on the 2% chart is the average yearly rate over the previous 20 years, which smooths wars and recessions but lags turning points.

Prediction verdicts

Too fast means the predicted change came later than the date named; too slow means it came sooner. Quotes are checked verbatim against the linked source, or shown as labelled paraphrases. Each open prediction states how it will be judged, and the daily sync flags any that pass their date for a verdict. Ambiguous is used when a claim is hedged or undefined.

Reader results

Quiz answers and prediction votes are pooled anonymously and shown only above a minimum sample (30 quiz takers, 10 votes per prediction). They describe this site's self-selected readers, not any population.

IT booms

A unit for comparing forecasts: about 0.67 percentage points a year of extra US labour productivity growth, the share of the late-1990s speed-up that Oliner and Sichel (2000) attribute to information technology. Comparisons across forecasts mix measures and are illustrations only.

Sources and licences

Sources marked “not synced yet” are used by sections still being built. Spotted an error? Tell me on LinkedIn. The forecast model and the weak-link framing are explained in the explainer, with the papers in the research library.